paulmweni/Predicts-Vaccine-Hesitancy-in-Africa

Machine Learning Models and Mixed Models to predict vaccine hesitancy in Africa # Vaccine Hesitancy in Africa: A Dual Machine Learning & Multilevel Analysis **Author:** Paul Marie Mweni | KU Leuven — Master's Thesis (QASS) --- ## Overview This project investigates the structural and individual-level determinants of **vaccine hesitancy** across **39 African countries** using data from the **Afrobarometer Round 9** (n ≈ 22,600). It combines machine learning for predictive modeling and feature evaluation with multilevel statistical modeling to account for the hierarchical, cross-national structure of the data. The dual-method approach bridges the gap between predictive power and causal interpretability — a combination increasingly valued in applied public health research and data science roles. --- ## Research Question > What are the strongest individual-level and country-level predictors of vaccine hesitancy among unvaccinated Africans, and how do these vary across national contexts? --- ## Methods ### Machine Learning Pipeline Three classifiers were benchmarked on the full sample and on two age-based subgroups (youth: 18–35; adults: 35+): | Model | Full Sample AUC | Youth AUC | Adult AUC | |----------|:-----------:|:---------:|:---------:| | Logistic Regression (GLM) | 0.629 | 0.624 | 0.632 | | Random Forest | 0.674 | 0.661 | 0.676 | | **XGBoost** | **0.686** | **0.674** | **0.691** | Key steps: - **Missing data imputation** via `missForest` (country-stratified, parallelized) - **Hyperparameter tuning** via 5-fold cross-validation with 30 grid combinations - **Optimal threshold selection** using the Youden Index - **Feature evaluation** with SHAP values (beeswarm + bar plots) and permutation importance ### Multilevel Analysis A stepwise **mixed-effects logistic regression** (GLMER, `lme4`) with countries as random intercepts: - **Null model ICC = 0.108** → 10.8% of variance attributable to country-level differences, justifying multilevel modeling - 10 nested models built incrementally: from language profile only → individual …